Developing an Uncertainty Aware Neural Network TGLF Surrogate Model for Fast Integrated Modeling of SPARC

To enable accelerated modeling of fusion devices recent research has focused on accelerating traditional codes with neural network (NN) trained surrogate models [Meneghini, NF 2017; Van de Plassche, PoP 2020; Alvarez, Manuscript Under Review 2024]. In this work we present the results of an uncertainty-aware TGLF NN that was developed for rapid prediction of the SPARC device parameter space. A training database was developed using SPARC primary reference discharge (PRD) [Creely, JPP 2020] and L-mode scenarios as its training parameter space. The NN utilizes 18 TGLF input parameters and is trained to predict heat and particle fluxes 3 (Q_e, Q_i, Γ_e). An uncertainty aware NN architecture was utilized to minimize the overall size of the training database and to highlight errors in the model that arises from training data that does not properly span the parameter space. Knowledge of this error was used to perform targeted parameter space training data generation. NNs that can predict transport fluxes with relative error compared to traditional TGLF simulations were developed. These models will be implemented into various profile prediction workflows such as ASTRA and PORTALS [Rodriguez-Fernandez, NF 2024] in order to compare the NN models with traditional profile prediction workflows with the ultimate goal of implementing the NN models into SPARC’s flight simulator framework.